Storm ECDSA Harness

Public repo: https://github.com/welttowelt/storm-ecdsa-harness

Fixture dashboard: https://welttowelt.github.io/storm-ecdsa-harness/

Storm is the public release of the STORM harness in production: an operator control plane for coordinating public ecdsa.fail research workflows. It is the layer around the solver: deciding what receives attention, forcing evidence labels, routing workers, gating compute, recording ACKs, and keeping public notes safe to share. The edge is multi-model collaboration: different workers can contribute through mailbox, API handoffs, or a named Zellij multiplexer board, but every claim still has to pass the same evidence, critique, and validation gates. It is not the original benchmark harness and it does not contain private hunt state.

Purpose and limits: ecdsa.fail is a public resource-estimation benchmark for the quantum cost of one secp256k1 point-add circuit. This repo contains workflow templates, evidence discipline, dashboard fixtures, and safety checks. It does not provide target selection, key recovery, private compute endpoints, raw scan logs, unreleased nonces, or live candidate diffs.

STORM Harness In Production

The live STORM controller is private. It refreshes the public leaderboard, reads the shared worker mailbox, creates route packets, dispatches CPU/GPU work, runs official validation, and opens the submit gate only after a fresh score check and explicit submit flag.

This public repo is the reusable release surface for that setup. It contains the protocol, templates, fixture dashboard, operator cards, and redaction checks so other OSS teams can copy the coordination layer without inheriting private state.

See docs/production-shape.md for the exact production-to-public mapping.

Zellij Control Plane Template

The current production pattern moved from loose manual terminal sessions plus a mailbox into one explicit Zellij board. The board gives the boss lane, policy surface, dispatch queue, wallbreakers, audit lanes, benchmark lane, provider governance lane, and long-context workers a stable place to live and resume.

Read docs/zellij-control-plane.md for the detailed model. Copy templates/zellij-control-plane/ into a private controller repo when you want a starter board with:

• layouts/fleet.kdl for the tab layout,
• bin/start-zellij and role launchers,
• policy and worker TSV examples,
• role goal prompts for boss, auditor, wallbreakers, Codex workers, and Claude long-context workers.

The template is intentionally public-safe. It contains no private worker hosts, provider account data, logs, raw nonces, live candidate state, or submit authority.

Operator Edge

Agentic benchmark work fails quietly when teams trust memory, chase stale routes, spend compute without a kill gate, or promote scanner hits into public claims too early. Storm makes that operating layer explicit:

• Current-target lock: every cycle starts from refreshed public benchmark state, not chat memory.
• Route packets: every idea names the claim, evidence level, validator, owner, budget, and falsifier before it receives compute.
• Multi-model workers: Codex-, Claude-, and DeepSeek-style workers can plug in as research leads, engineering specialists, deep researchers, or external pressure-test reviewers.
• Critique loop: constructive reviews, skeptical reviews, source checks, and smallest-useful-fix passes run before compute or public claims.
• Compute gate: CPU/GPU work is allowed only when a route has a predicate, validator, owner, budget, and stop condition.
• Submit gate: no win language until fresh frontier recheck, official local validation, score win, legal diff, public note, and explicit submit flag.
• Public boundary: dashboards, examples, and notes stay fixture/public-only while private logs, endpoints, routes, and nonces stay out.

The repo is therefore not a faster circuit generator by itself. It is a way to make a group of agents and humans behave like a disciplined research desk.

Lineage

• Base challenge / benchmark lineage: Gautham Anant and the public ecdsa.fail challenge. The public framing is described in Sreeram Kannan's article: https://www.linkedin.com/pulse/you-beat-google-sreeram-kannan-rqehc.
• Early custom agent-harness inspiration: Gajesh Naik / Gajesh2007, including custom agent workflows, goal-mode style operation, model/subagent routing, and more agent-readable code organization.
• Storm contribution: the STORM harness control-plane layer: mailbox protocol, ACK/read receipts, route packets, worker roles, dashboard views, compute gates, submit gates, and redaction discipline.

See docs/credits.md and CREDITS.yaml for explicit attribution. Credits are split into direct tools, public sources, and group-discussion process inspiration so the repo does not imply copied code or endorsement where there was only inspiration.

What This Repo Contains

• A current-target-first operating model.
• A production-shape map for the live STORM controller.
• Worker role cards and handoff templates.
• Public operator cards under operators/ for Kimi Storm, Deep Storm, and Codex Storm handoffs.
• Historical route packets under routes/ with falsification status where known.
• Critique-before-compute audit loop.
• Evidence labels and validation gates.
• Route packet and compute request templates.
• External pressure-test worker prompt pattern.
• Safe filled examples for the templates.
• A fixture-only dashboard for operator status.
• A redaction checklist and automated redaction scan.
• Machine-readable claim-ledger, proof-backlog, pod-admission, and state-packet helper scripts for running short control loops without private state.
• Public carry-toy gates that separate value/recompute hypotheses from compute or submit authorization.
• Public patch packages for source-preserving throughput helpers that require their own validation gates before use.
• Public-credit policy for community-derived ideas.
• Repo-local Codex skill bridges for ECDSA workflow agents.
• A sanitized Zellij multiplexer control-plane template for replacing manual session sprawl with named tabs, visible gates, dispatch state, and role goals.

Skill Cards

The skills/ directory contains prompt cards that can be copied into agent instructions or used as a preflight before route work, diagnosis, and verification:

• skills/nasqret-playbook.md: build route slates before scaling compute.
• skills/deepseek-pressure-test.md: ask an external worker to try to kill a route before acting.
• skills/tony-rci-audit.md: turn critique into concrete findings, smallest useful fixes, and explicit gates.
• skills/bluesky-audit.md: find the best bounded path if an idea is real before deciding whether to invest.
• skills/redsky-audit.md: try to kill a route with stale-truth, evidence, validator, economics, and gate checks.
• scripts/storm-bluesky-redsky-loop-gate.py: validate a five-loop Bluesky/Redsky process packet with implemented findings and no compute/submit overclaim.
• skills/agent-unstuck.md: get a stuck agent moving again by diagnosing the real failure, testing a different explanation, and verifying the result across software, operations, research, writing, analysis, and business development. The page is self-contained and can be given directly to an agent by URL; no repository clone is required.
• skills/pip-discipline.md: use the English/PIP-derived evidence discipline for ECDSA-specific public coordination after repeated failure, passive handoff, or unverified completion claims.
• skills/exact-support-miner.md: turn public op-trace facts into ranked value-exact skip packets with proof and validation gates.
• skills/frontier-lock.md: refresh and record the current public frontier before route, compute, note, or submit decisions.
• skills/validation-submit-gate.md: keep evidence labels, trusted validation, and submit-gate claims separate.
• skills/route-compute-gate.md: require owner, predicate, validator, budget, and stop condition before CPU/GPU dispatch.
• skills/multi-agent-handoff.md: make cross-worker handoffs explicit with ACKs, validators, and kill gates.
• skills/benchmark-frontier-archaeology.md: walk adjacent public frontier submissions to extract transferable levers, reject search artifacts, and name the next binder.
• skills/circuit-diff-mining.md: mine recent promoted diffs into bounded candidate probes without stale knob copying.
• skills/circuit-optimization-intake.md: turn route ideas into ranked, source-bound packets with score estimates, falsifiers, and repair-before-hunt gates.
• skills/peak-qubit-reduction.md: inventory live-set binders, co-binder plateaus, and qubit-Toffoli tradeoffs before structural width cuts.
• skills/toffoli-reduction.md: classify emitted/executed Toffoli hot spots and gate count-cut primitives before promoting a Toffoli route.
• skills/reversible-circuit-validation.md: validate candidate routes across classical, phase, and ancilla channels before clean or submit language.
• skills/island-hunting-bakeoff.md: compare score-eligible routes with equal-size pilot scans before committing large GPU ranges.
• skills/submission-forensics.md: deep-audit public submissions, notes, promoted diffs, and adjacent frontier moves without private-state leakage or credit overclaiming.
• skills/bluesky-route-salvage.md: find the smallest falsifiable experiment that preserves upside before a route is killed.
• skills/redsky-frontier-audit.md: run adversarial frontier, legality, evidence-label, compute, and submit-gate checks.
• skills/ecdsafail-cli-ops.md: operate the ecdsafail CLI with worktree, credential, sync/reset, validation, and submit discipline.
• skills/stormgate-prefilter.md: keep prefilter, canary, survivor-label, and trusted-validation boundaries explicit.
• skills/q1152-structural-core.md: gate lower-q structural routes on global wall coverage, proof, score economics, count screens, and trusted eval.
• skills/construction-package-gate.md: reject lower-q construction packages that lack q-tier score edge, co-binder coverage, or clean validation status.
• skills/construction-intake-gate.md: require paper-mined construction packets to name a current source replacement, source/candidate hashes, q/T economics, restore/phase/ancilla obligations, and a bounded toy falsifier before review.
• skills/anvil-mass-ledger-gate.md: require conditional-Toffoli mass ledgers to be machine-readable, source-bound, score-consistent, and routed to the right next gate before workers promote rows.
• skills/anvil-namespace-gate.md: require high-mass Anvil op_index rows to declare the index namespace and source binding before exact-support proof work.
• skills/anvil-exact-redundancy-gate.md: require one high-mass Anvil op to carry exact-support, restore, phase, ancilla, allocator, and score evidence before review.
• skills/pebbling-theorem-gate.md: require recompute/pebbling packets to name the DAG node, producer/consumers, recompute path, score edge, and certified restore/phase/ancilla/support proofs before route review.
• skills/frontier-escape-gate.md: classify post-local-optimum packets as source theorem, construction package, nonce retune, or NACK.
• skills/single-ccx-fanout-throughput.md: apply and verify the d44 SINGLE_CCX_FANOUT no-clone throughput patch without promoting it into a clean-route claim.
• skills/fanout-runpod-qstate-guard.md: verify fanout GPU qstate headers and hashes before spending pod time.
• skills/fanout-survivor-phase-gate.md: gate fanout GPU survivors against phase-aware official eval evidence before any win-language handoff.
• skills/official-fast-exit-eval.md: apply and use the eval_circuit ISLAND_FAST_EXIT dirty-triage patch without turning partial dirty counts into score claims.
• skills/official-eval-isolation-gate.md: require lock or isolated workdir evidence before remote official eval output can become submit-grade evidence.
• skills/fleet-owner-claim-gate.md: require owner, pod identity, route/range, active watcher/log evidence, next action, and no_submit_ack=yes before a paid instance survives audit.
• skills/local-heavy-compute-gate.md: block Mac-local heavy ECDSA compute and recurring watcher wrappers from laptop process snapshots, including storm-exact-miner, route_compare, search_driver, lower-q, proof, and scanner loops.
• skills/pod-inventory-ack-gate.md: accept complete provider inventory ACKs, hold unknown/unreachable inventory, accept verified zero-pod account inventory, and fail ownerless running pods, nonzero empty-account spend, or compute start language after fleet stop directives.
• skills/route-compare-admission-gate.md: require clean baseline, clean candidate, clean compare output, and a strict score edge before route-compare evidence can become residual, compute, pod, handoff, submit, or alert work.
• skills/ffg-pair-proof-gate.md: require pair-complete FFG call coverage, certified value/restore/phase proof, source and candidate hashes, route admission, sufficient shot depth, positive score edge, and no-submit discipline before any FFG suffix-carry handoff or compute language.
• skills/compute-restart-gate.md: block scanner restarts when compute is closed, no route ACK exists, or no certified source/full-clean evidence exists.
• skills/compute-unlock-gate.md: require a source-hash-bound certified value-exact packet, exact diff, negative edge, allocator order, validation owner, budget, stop condition, Storm route ACK, and no-submit ACK before compute can reopen.
• skills/candidate-validation-packet-gate.md: check remote validation packets for host, owner, ops, score edge, c/p/a, artifacts, and no-submit discipline before FOR-AKASH language.
• skills/source-packet-novelty-gate.md: admit only source-bound, candidate hash-bound, negative-edge packets outside the closed ledger; fail exhausted source-family summaries before any proof or compute request.
• skills/source-row-routing-gate.md: require each q1152 scout row to become either one packet-ready bounded-proof row or one counterexample closure that advances to the next source row, with no compute authority.
• skills/gidney-ccz-residual-gate.md: screen Gidney erase-CCZ residual packets so default-on remainder rows and below-bar mass do not unlock compute.
• skills/qoffset-host-accounting-gate.md: require source-integrated qoffset packets or closures to carry source/candidate hashes, proof status, restore and phase flags, counted q1152 delta, and no-submit discipline.
• skills/emit-bundle-support-gate.md: require combined real-emit bundles to prove one shared invariant, clear the same-q shot-mass bar, and close cleanly when witnesses differ or rows are live.
• skills/transcript-overlap-gate.md: gate transcript peak-overlap drop packets before implementation, blocking active-only origins, stale indexes, dirty probes, missing peak calls, and premature residual/compute requests.
• skills/paper-gidney-constant-workspace-adder.md: apply Gidney 2025 constant-workspace adders only after a toy proof and one-callsite residual gate.
• skills/paper-mbu-modular-arithmetic.md: turn measured-workspace arithmetic cleanup into an explicit phase-correction proof obligation.
• skills/paper-hrs-dirty-constant-adder.md: use HRS dirty-ancilla adders as a disciplined borrowed-host route, not as zero scratch.
• skills/paper-gidney-temporary-logical-and.md: evaluate temporary logical-AND erase patterns for uncompute and Toffoli budgeting.
• skills/paper-haner-ecdlp-circuits.md: compare point-add route ideas against established ECDLP Q/T/depth tradeoffs.
• skills/paper-schrottenloher-point-addition.md: map 2026 secp256k1 point-addition architecture ideas into local TLM invariants.
• skills/paper-schrottenloher-dialog-codec-audit.md: test Schrottenloher dialog-codec compression ideas against entropy bounds before lower-Q transcript edits.
• skills/paper-luo-register-sharing-eea.md: explore compact register-sharing EEA ideas without hiding decompression peaks.
• skills/conditionally-clean-cascade-cut.md: test the Nie-Zi-Sun / Khattar-Gidney conditionally-clean cascade route for exact peak-carry cuts.
• skills/paper-conditionally-clean-ancillae.md: use Khattar/Gidney conditionally clean ancillae only with a branch-clean proof.
• skills/paper-reversible-pebbling-memory-management.md: turn lower-Q cleanup/recompute ideas into explicit peak DAG and stale-drop gates.
• skills/paper-square-active-volume.md: rank reuse, uncompute, and codec timing candidates by the above-target active-volume rows they actually cover.
• skills/paper-vandaele-optimal-comparator.md: apply Vandaele 2026 comparator/incrementer circuits only after proving the comparator is the peak source or pairing it with the co-peak carry floor.
• skills/paper-remaud-ancilla-free-adder.md: evaluate no-ancilla adder suffixes as lower-Q pressure relief with a strict Toffoli gate.
• skills/paper-takahashi-no-ancilla-adder.md: use classic no-ancilla addition as a suffix-only fallback baseline.
• skills/paper-invariant-intake-gate.md: block paper-only scout hits and admit only source-backed, candidate-hash-bound, certified local invariants to skill-card review, without compute authority.
• skills/paper-roetteler-ecdlp-resource-estimate.md: ground prime-field ECDLP architecture ideas in gate-level point-add validation.
• skills/paper-garn-kan-windowed-binary-ecdlp.md: borrow exact exceptional-case and window/table discipline without direct field-porting.
• skills/paper-wire-recycling-lifetime-graph.md: apply wire-recycling, SQUARE, and Reqomp lifetime-graph discipline to free-pool aliasing and stale-dead-drop failures.
• skills/paper-dirty-borrowing-entanglement.md: require idle-source, original-state, and external-entanglement identity proofs before borrowed working qubits can support lower-Q cuts.
• skills/paper-scalable-memory-recycling.md: prevent lower-Q primitive swaps from silently changing dynamic headroom schedules; require schedule ledgers, frozen baseline replay, and explicit rebaked tables.
• skills/paper-dead-gate-elimination.md: regenerate and audit dead-drop index files after op-stream edits; never reuse stale .idx proof artifacts.

Codex-style agents can also discover the general Agent Unstuck card, the ECDSA-specific PIP card, and the audit cards through the repo-local bridges under .agents/skills/.

How To Call The Skills

In Codex or another skill-aware agent, call a repo-local skill by name:

Use frontier-lock.
Use bluesky-route-salvage on this route.
Use redsky-frontier-audit before compute.
Use ecdsafail-cli-ops and validation-submit-gate before submit.

To use Agent Unstuck without cloning this repository, send the agent this URL:

https://github.com/welttowelt/storm-ecdsa-harness/blob/main/skills/agent-unstuck.md

Then say: Read the entire skill at this link and apply it to the current task.

The bridge names are:

• nasqret-playbook
• deepseek-pressure-test
• agent-unstuck
• pip-discipline
• tony-rci-audit
• exact-support-miner
• frontier-lock
• validation-submit-gate
• route-compute-gate
• multi-agent-handoff
• benchmark-frontier-archaeology
• circuit-diff-mining
• circuit-optimization-intake
• peak-qubit-reduction
• toffoli-reduction
• reversible-circuit-validation
• island-hunting-bakeoff
• submission-forensics
• bluesky-route-salvage
• redsky-frontier-audit
• ecdsafail-cli-ops
• stormgate-prefilter
• fanout-survivor-phase-gate
• official-fast-exit-eval
• official-eval-isolation-gate
• fleet-owner-claim-gate
• local-heavy-compute-gate
• route-compare-admission-gate
• compute-restart-gate
• compute-unlock-gate
• candidate-validation-packet-gate
• qoffset-host-accounting-gate
• emit-bundle-support-gate
• construction-intake-gate
• anvil-mass-ledger-gate
• anvil-namespace-gate
• anvil-exact-redundancy-gate
• pebbling-theorem-gate
• transcript-overlap-gate
• paper-gidney-constant-workspace-adder
• paper-mbu-modular-arithmetic
• paper-hrs-dirty-constant-adder
• paper-gidney-temporary-logical-and
• paper-haner-ecdlp-circuits
• paper-schrottenloher-point-addition
• paper-schrottenloher-dialog-codec-audit
• paper-luo-register-sharing-eea
• conditionally-clean-cascade-cut
• paper-conditionally-clean-ancillae
• paper-reversible-pebbling-memory-management
• paper-vandaele-optimal-comparator
• paper-remaud-ancilla-free-adder
• paper-takahashi-no-ancilla-adder
• paper-roetteler-ecdlp-resource-estimate
• paper-garn-kan-windowed-binary-ecdlp
• paper-wire-recycling-lifetime-graph
• paper-dirty-borrowing-entanglement
• paper-scalable-memory-recycling
• paper-dead-gate-elimination

Useful global skills for ECDSA work, when installed in the agent environment:

• redsky: deep ECDSA.fail frontier-solving audit.
• ecdsafail-cli: detailed CLI syntax for benchmark, run, submit, notes, sync, reset, and skill install.

Common call chains:

• General company work after repeated failure or an unverified completion claim: agent-unstuck; add tony-rci-audit when revising an existing artifact.
• Full route-intake stack: frontier-lock -> benchmark-frontier-archaeology -> circuit-optimization-intake -> peak-qubit-reduction / toffoli-reduction -> reversible-circuit-validation -> island-hunting-bakeoff -> route-compute-gate -> validation-submit-gate.
• Exact skip mining: frontier-lock -> exact-support-miner -> redsky-frontier-audit -> validation-submit-gate.
• Frontier shift: frontier-lock -> nasqret-playbook -> benchmark-frontier-archaeology -> circuit-diff-mining.
• New route: frontier-lock -> bluesky-route-salvage -> redsky-frontier-audit -> route-compute-gate.
• External critique: deepseek-pressure-test -> redsky-frontier-audit.
• Public submission audit: frontier-lock -> submission-forensics -> circuit-diff-mining.
• Prefilter/search work: frontier-lock -> stormgate-prefilter -> route-compute-gate -> validation-submit-gate.
• Submit attempt: frontier-lock -> ecdsafail-cli-ops -> validation-submit-gate.
• Worker handoff: multi-agent-handoff plus whichever route or validation card owns the task.
• Paper-driven lower-q adder work: paper-gidney-constant-workspace-adder -> paper-mbu-modular-arithmetic -> q1152-structural-core -> construction-package-gate.
• Post-local-optimum route triage: frontier-lock -> frontier-escape-gate -> validation-submit-gate.
• Conditionally-clean cascade cut: conditionally-clean-cascade-cut -> paper-conditionally-clean-ancillae -> fold-carry-rearchitecture -> exact-support-invariant-miner.
• Free-pool or stale-dead-drop integration failure: paper-wire-recycling-lifetime-graph -> redsky-frontier-audit -> validation-submit-gate.
• Borrowed working-lane cut: paper-dirty-borrowing-entanglement -> paper-wire-recycling-lifetime-graph -> paper-dead-gate-elimination.
• Schedule-coupled lower-Q primitive: paper-scalable-memory-recycling -> frontier-lock -> structural-qubit-cut -> validation-submit-gate.
• Comparator peak pressure: paper-vandaele-optimal-comparator -> q1152-structural-core -> exact-support-invariant-miner.
• Peak host borrowing: paper-conditionally-clean-ancillae -> fold-carry-rearchitecture -> exact-support-invariant-miner.
• Zero-ancilla suffix fallback: paper-remaud-ancilla-free-adder -> paper-takahashi-no-ancilla-adder -> route-compute-gate.
• Paper-driven architecture work: paper-haner-ecdlp-circuits -> paper-schrottenloher-point-addition -> paper-luo-register-sharing-eea -> paper-roetteler-ecdlp-resource-estimate, then reduce the idea to one local invariant.

Quickstart

Use this as a GitHub template or clone it directly:

git clone https://github.com/welttowelt/storm-ecdsa-harness.git
cd storm-ecdsa-harness
scripts/redaction-check.sh
scripts/redaction-check.sh --history
scripts/check-public-harness.sh
python3 -m http.server 8787 --directory dashboard

Then adapt the templates with fixture or public data only. Keep private hunt state in a separate private repo or local folder.

Full guide: docs/quickstart.md.

Safe filled examples: examples/.

Audit loop: docs/audit-loop.md.

Prefilter gate pattern: docs/stormgate.md.

Latest stormgate Redsky audit: docs/redsky-stormgate-audit-2026-06-20.md.

What This Repo Does Not Contain

• Live ecdsafail-hunt-ops state.
• Private mailbox history.
• Real run logs, pod endpoints, keys, tokens, or account details.
• Unreleased candidate diffs or winning nonces.
• Active strategy notes from an ongoing hunt.

Operating Model

Every cycle starts from public truth, not memory:

1. Refresh the live benchmark/current target.
2. Read mailbox entries and ACK directed handoffs.
3. Check route packets and evidence labels.
4. Run cheap validation before paid compute.
5. Dispatch compute only when a route has a predicate, owner, validator, kill gate, and bounded stop condition.
6. Treat scanner hits as Prefilter until official local validation passes.
7. Submit only after fresh frontier recheck, official local 0/0/0, score win, legal narrow diff, public note, and explicit submit flag.

More detail: docs/operating-model.md.

Dashboard

The dashboard in dashboard/ is static and fixture-only. It shows how to surface current target, worker status, mailbox activity, breaking news, fleet state, and submit gates without leaking operational details.

Public fixture dashboard:

https://welttowelt.github.io/storm-ecdsa-harness/

Open locally:

python3 -m http.server 8787 --directory dashboard

Then open the local server URL printed by your terminal.

Redaction Check

Run this before publishing or pushing:

scripts/redaction-check.sh

The check rejects common private material: API-key shapes, remote command fragments, host/port patterns, private paths, raw nonce assignments, live mailbox filenames, and forbidden state directories.

Before a public push, also scan committed history:

scripts/redaction-check.sh --history

The GitHub Actions workflow runs both checks on push and pull request.

Release Rule

Do not publish until:

• scripts/redaction-check.sh passes.
• scripts/redaction-check.sh --history passes.
• docs/credits.md names every borrowed idea or tool.
• All examples use fixture/demo data.
• A human review confirms no private hunt state remains.

See docs/publish-checklist.md for the full publish sequence and docs/telegram-announcement.md for a short sharing draft.